Fix potential autocast NaNs in image upscale

[Cyberbeing]

Description

Bugfix for potential autocast NaNs in image upscale split off from Pull Request #14794

Discussion in that other pull request now confirms this works for others for preventing the black screen NaNs with Float32 upscaler model architectures [1] [2]. The performance impacts for Float16 converted models was also recently confirmed to be minimal after some recent changes [3].

Float32 upscaler models may now be slower, since they will be processed more than before in Float32 instead of autocast, but this seems necessary to fix the NaNs. In the future, I imagine this performance impact could be lifted by supporting BFloat16 upscaling, which Spandrel seems to support for every model architecture unlike Float16.

Spandrel Arch supporting Float32/BFloat16 only: CodeFormer, DAT, GFPGAN, GRL, HAT, SRFormer, SwinIR

Spandrel Arch supporting Float32/Float16/BFloat16: Compact, ESRGAN, RealESRGAN, OmniSR, SPAN
Note: Float16 support for ESRGAN and others requires the additional changes in #14794. Currently of the currently supported model types, prefer_half which triggers float16 model conversion only seems to be enabled for realesrgan_model.py but not esrgan_model.py

• • Fixed an autocast NaN during upscaling I've noticed rarely on my system during Hires fix as mentioned in:
    [Bug]: DAT upscaler is not working in hires fix. #14710 (comment)
    Support more Upscalers #14794

Checklist:

• I have read contributing wiki page
• I have performed a self-review of my own code
• My code follows the style guidelines
• My code passes tests